Detecting abbreviations in discharge summaries using machine learning methods.

Area of Science:

• Clinical Natural Language Processing
• Medical Informatics
• Machine Learning Applications

Background:

• Accurate recognition of clinical abbreviations is crucial for effective clinical natural language processing (NLP).
• A comprehensive lexical resource of clinical abbreviations is needed for improved NLP applications.
• Current methods for abbreviation identification may lack sufficient accuracy and scalability.

Purpose of the Study:

• To develop and evaluate a corpus-based machine learning method for creating a lexical resource of clinical abbreviations.
• To automatically detect clinical abbreviations from hospital discharge summaries.
• To assess the performance of various machine learning algorithms in clinical abbreviation identification.

Main Methods:

• A corpus of seventy hospital discharge summaries was manually annotated by domain experts.
• The annotated data was split into training (40 documents) and testing (30 documents) sets.
• Several machine learning algorithms, including Random Forest, were implemented and evaluated using pre-defined features.

Main Results:

• The Random Forest classifier achieved a high F-measure of 94.8% (98.8% precision, 91.2% recall) on the test set.
• A voting scheme combining multiple machine learning classifiers further improved performance, reaching an F-measure of 95.7%.
• The developed system demonstrated strong capabilities in automatically detecting clinical abbreviations.

Conclusions:

• The corpus-based machine learning approach is effective for building a clinical abbreviation lexical resource.
• Machine learning models, particularly Random Forest and ensemble methods, show high accuracy in identifying clinical abbreviations.
• This work contributes to advancing clinical NLP by providing a robust tool for abbreviation recognition.